Sushko S. Methods of optimal parallelization of programs of microprocessor systems to increase its efficiency

Українська версія

Thesis for the degree of Candidate of Sciences (CSc)

State registration number

0421U102748

Applicant for

Specialization

  • 05.13.05 - Комп'ютерні системи та компоненти

14-05-2021

Specialized Academic Board

Д 26.185.01

Institute of Problems of Modeling in Energy named after GE Pukhov of the National Academy of Sciences of Ukraine

Essay

In the Ph.D. thesis new scientific and applied results to increase of execution efficiency of software by using of tiling method with choice of the best sizes of tiles are received. The obtained experimental data indicate an acceleration of processing and energy efficiency of calculations for most of test programs by using of the proposed method, named Smart Tiling Method. The object of research is a process of software optimization of loops’ operators of microprocessor programs. The aim of the work is to increase of efficiency of microprocessor systems through parallelization and optimization of software. The subject of research is complexes of automatic software development of microprocessor systems. Research methods are based on set theory, computational loop transformation methods and evolutionary optimization methods. The novelty of scientific statements and results obtained personally by the applicant is as follows: Smart Tiling Method was developed which searches for optimal dimensions of rectangular tiles of iterative space of microprocessor loop operators. The general approach of parallelization of loop operators of microprocessor programs is proposed, which is aimed to find an optimal solution for dividing of iterative space of loop operator into tiles. The comprehensive assessment of parallelization/optimization quality of applied software was proposed which is based on energy efficiency of calculations. The method of estimation of execution time of test programs was improved by using of variety of tiling methods, which is based on automatic sequential multiple runs of test programs and storing of results which can be further analyzed. Discrete Particle Swarm Optimization Method was improved by determining of method’s coefficients and by determining of best initial particles’ positions. Results of the work are implemented in LLC “Delta SPE”. Results can be applied in optimization of computer programs written in the C or C++ programming language to accelerate their execution time. Smart Tiling Method is universal in terms of program’s purpose to be accelerated, hardware platform and used compiler.

Files

Similar theses